Deep Learning Based Unsupervised Biomedical Image Restoration

Biomedical image is easily polluted by noise or artifact in the process of acquisition,which may affect the image quality and even the correct diagnosis of the disease.Biomedical image restoration is an important research content in the field of image processing,including medical image denoising,artifact removal,image reconstruction and other tasks,which has high application value in the actual biomedical imaging.This paper mainly deals with three kinds of biomedical images,including intravascular ultrasound image,computed tomography(CT)image and fluorescence microscopic image.Among them,intravascular ultrasound image and CT image belong to medical imaging and fluorescence microscopic image belongs to biological image.Each biomedical image has different degradation models(the principle of artifact and noise).For intravascular ultrasound image,the guide wire artifact is a common image degradation model,which is produced by the metal guide wire occlusion ultrasound in the imaging system.For CT images,low-dose CT images are used to reduce the radiation injury of patients.Due to the reduction of radiation dose,the quantum noise in the image acquisition process is caused.This quantum noise presents Poisson distribution.For fluorescence microscopic images,the degradation model is Poisson Gaussian noise,which is caused by the physical limitations of the components in the fluorescence microscopic system,including Poisson noise caused by the micro image detector and the electronic thermal disturbance,as well as the additive Gaussian noise caused by the electronic reading equipment.The biomedical image restoration tasks involved in this paper include the removal of intravascular ultrasound guide wire artifacts,low-dose CT images and fluorescence microscopic image denoising.With the rapid development of deep learning technology,the convolutional neural network model has become the mainstream method of biomedical image restoration.The training of these deep learning models requires two important conditions:(1)huge data volume;(2)pair of noisy and noiseless images need to be supervised.This kind of supervised deep learning model is difficult to apply to the actual biomedical environment,because:(1)it is difficult to obtain the paired noisy and noiseless biomedical images;(2)the cost of biomedical image acquisition is high;(3)the amount of biomedical image data is generally small.In order to solve the shortcomings of the current depth learning algorithm,this paper proposes two unsupervised depth learning algorithms,which are respectively applied to the tasks of removing the artifacts of intravascular ultrasound guide wire,low-dose CT image and fluorescence microscopic image denoising.For the task of removing the artifact of intravascular ultrasound guide wire,this paper proposes an unsupervised deep learning method based on attention guidance.The proposed approach only need unpaired data.The main innovations are:(1)using the attention mechanism to make the model only focus on the guide wire artifact area in training,while keeping other normal imaging areas unchanged;(2)using the useful information between adjacent frames of intravascular ultrasound image sequence to improve the effect of artifact removal.For the denoising task of low dose CT image and fluorescence microscopic image,this paper proposes a deep learning method based on self supervised learning.The main idea is to use the correlation between pixels in biomedical image to achieve the purpose of denoising.This method only needs noise image to be trained.At the same time,this method can be trained and tested on a single image without building a training set.